Skip to content

GitLab

"profiler/vscode:/vscode.git/clone" did not exist on "8ebea3a56e68b6b638c78f3644b53f3c20e245f2"
Commit f3bbfe3e authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of https://github.com/ROCm/composable_kernel into update_cka8w8

parents 2b840f5a efb34741
Hide whitespace changes
Inline Side-by-side
Showing with 1104 additions and 508 deletions
include/ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp 0 → 100644
View file @ f3bbfe3e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/softmax.hpp"
#include "ck_tile/ops/topk.hpp"
namespace ck_tile {
struct MoeSortingPolicy
{
};
} // namespace ck_tile
include/ck_tile/ops/fused_moe/pipeline/moe_sorting_problem.hpp 0 → 100644
View file @ f3bbfe3e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
#include <string>
#include <type_traits>
namespace ck_tile {
template <typename IndexType_, typename WeightType_, index_t InternalLoadUnroll_>
struct MoeSortingProblem
{
// TODO: this kernel only support warp per row
using WeightType = remove_cvref_t<WeightType_>;
using IndexType = remove_cvref_t<IndexType_>;
static constexpr index_t WarpSize = get_warp_size();
static constexpr index_t WarpsPerBlock = 1;
static constexpr index_t InternalLoadUnroll = InternalLoadUnroll_;
};
} // namespace ck_tile
include/ck_tile/ops/gemm/kernel/gemm_kernel.hpp
View file @ f3bbfe3e
......@@ -115,12 +115,22 @@ struct GemmKernel
}
}();
auto a_pad_view = pad_tensor_view(
a_tensor_view,
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
// somehow clang-format is splitting below line into multiple.
// clang-format off
sequence<false, GemmPipeline::kPadA>{});
auto a_pad_view = [&]() {
if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
{
return pad_tensor_view(
a_tensor_view,
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
sequence<false, GemmPipeline::kPadK>{});
}
else
{
return pad_tensor_view(
a_tensor_view,
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
sequence<GemmPipeline::kPadM, false>{});
}
}();
// clang-format on
auto a_block_window = make_tile_window(
......@@ -128,12 +138,22 @@ struct GemmKernel
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
{i_m, 0});
auto b_pad_view = pad_tensor_view(
b_tensor_view,
make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}),
// clang-format off
sequence<false, GemmPipeline::kPadB>{});
// clang-format on
auto b_pad_view = [&]() {
if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
{
return pad_tensor_view(
b_tensor_view,
make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}),
sequence<false, GemmPipeline::kPadK>{});
}
else
{
return pad_tensor_view(
b_tensor_view,
make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}),
sequence<GemmPipeline::kPadN, false>{});
}
}();
auto b_block_window = make_tile_window(
b_pad_view,
......@@ -171,18 +191,28 @@ struct GemmKernel
}
}();
auto c_pad_view = pad_tensor_view(
c_tensor_view,
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
// clang-format off
sequence<false, GemmPipeline::kPadC>{});
// clang-format on
auto c_block_window = make_tile_window(
auto c_pad_view = [&]() {
if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
{
return pad_tensor_view(
c_tensor_view,
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
sequence<false, GemmPipeline::kPadN>{});
}
else
{
return pad_tensor_view(
c_tensor_view,
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
sequence<GemmPipeline::kPadM, false>{});
}
}();
auto CBlockWindow_pad = make_tile_window(
c_pad_view,
make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
{i_m, i_n});
EpiloguePipeline{}(c_block_window, c_block_tile);
EpiloguePipeline{}(CBlockWindow_pad, c_block_tile);
}
};
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp
View file @ f3bbfe3e
......@@ -113,9 +113,9 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
static constexpr index_t VectorSizeB = Problem::VectorSizeB;
static constexpr index_t VectorSizeC = Problem::VectorSizeC;
static constexpr bool kPadA = Problem::kPadA;
static constexpr bool kPadB = Problem::kPadB;
static constexpr bool kPadC = Problem::kPadC;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadK = Problem::kPadK;
// Where is the right place for HasHotLoop and TailNum ???
static constexpr bool HasHotLoop = Problem::HasHotLoop;
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1.hpp
View file @ f3bbfe3e
......@@ -33,9 +33,9 @@ struct GemmPipelineAGmemBGmemCRegV1
static constexpr index_t VectorSizeB = Problem::VectorSizeB;
static constexpr index_t VectorSizeC = Problem::VectorSizeC;
static constexpr bool kPadA = Problem::kPadA;
static constexpr bool kPadB = Problem::kPadB;
static constexpr bool kPadC = Problem::kPadC;
static constexpr bool kPadM = Problem::kPadM;
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadK = Problem::kPadK;
CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize()
{
......@@ -101,11 +101,8 @@ struct GemmPipelineAGmemBGmemCRegV1
Policy::template MakeADramTileDistribution<Problem>());
// A LDS tile window for store
auto a_copy_lds_window =
make_tile_window(a_lds_block,
make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}),
{0, 0},
a_copy_dram_window.get_tile_distribution());
auto a_copy_lds_window = make_tile_window(
a_lds_block, make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}), {0, 0});
// B DRAM tile window for load
auto b_copy_dram_window =
......@@ -115,11 +112,8 @@ struct GemmPipelineAGmemBGmemCRegV1
Policy::template MakeBDramTileDistribution<Problem>());
// B LDS tile window for store
auto b_copy_lds_window =
make_tile_window(b_lds_block,
make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}),
{0, 0},
b_copy_dram_window.get_tile_distribution());
auto b_copy_lds_window = make_tile_window(
b_lds_block, make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}), {0, 0});
// A LDS tile for block GEMM
auto a_lds_gemm_window = make_tile_window(
......@@ -149,12 +143,32 @@ struct GemmPipelineAGmemBGmemCRegV1
tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
// LDS write 0
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
store_tile(a_copy_lds_window, a_block_tile_tmp);
if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::ColumnMajor>)
{
auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
Policy::template MakeShuffledARegBlockDescriptor<Problem>());
shuffle_tile(a_shuffle_tmp, a_block_tile);
const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_shuffle_tmp);
store_tile(a_copy_lds_window, a_block_tile_tmp);
}
else
{
store_tile(a_copy_lds_window, tile_elementwise_in(a_element_func, a_block_tile));
}
// LDS write 0
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
store_tile(b_copy_lds_window, b_block_tile_tmp);
if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
{
auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegBlockDescriptor<Problem>());
shuffle_tile(b_shuffle_tmp, b_block_tile);
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_shuffle_tmp);
store_tile(b_copy_lds_window, b_block_tile_tmp);
}
else
{
store_tile(b_copy_lds_window, tile_elementwise_in(b_element_func, b_block_tile));
}
}
index_t iCounter = num_loop - 1;
......@@ -180,8 +194,19 @@ struct GemmPipelineAGmemBGmemCRegV1
store_tile(a_copy_lds_window, a_block_tile_tmp);
// LDS write i + 1
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
store_tile(b_copy_lds_window, b_block_tile_tmp);
if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
{
auto b_shuffle_tmp_loop = make_static_distributed_tensor<BDataType>(
Policy::template MakeShuffledBRegBlockDescriptor<Problem>());
shuffle_tile(b_shuffle_tmp_loop, b_block_tile);
store_tile(b_copy_lds_window,
tile_elementwise_in(b_element_func, b_shuffle_tmp_loop));
}
else
{
const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
store_tile(b_copy_lds_window, b_block_tile_tmp);
}
iCounter--;
}
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp
View file @ f3bbfe3e
......@@ -11,6 +11,7 @@ namespace ck_tile {
// Default policy class should not be templated, put template on member functions instead
struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
{
#if 0
// 2d
template <typename Problem>
......@@ -116,6 +117,20 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
return smem_size;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackA()
{
using ADataType = remove_cvref_t<typename Problem::ADataType>;
return Problem::VectorLoadSize / sizeof(ADataType);
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackB()
{
using BDataType = remove_cvref_t<typename Problem::BDataType>;
return Problem::VectorLoadSize / sizeof(BDataType);
}
#elif 1
// fake XOR
template <typename Problem>
......@@ -192,80 +207,269 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
{
using ADataType = remove_cvref_t<typename Problem::ADataType>;
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t K1 = 16 / sizeof(ADataType);
constexpr index_t K0 = kKPerBlock / K1;
constexpr index_t M2 = get_warp_size() / K0;
#if 1 // coalesce reading for each blocks
constexpr index_t M1 = kBlockSize / get_warp_size();
static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
constexpr index_t M0 = kMPerBlock / (M2 * M1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
#else // coalesce reading for each warps
constexpr index_t M0 = kBlockSize / get_warp_size();
constexpr index_t M1 = kMPerBlock / (M2 * M0);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<0>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
#endif
using ALayout = remove_cvref_t<typename Problem::ALayout>;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
constexpr index_t M1 = Problem::VectorLoadSize / sizeof(ADataType);
constexpr index_t M0 = MPerBlock / M1;
constexpr index_t total_pixels = MPerBlock * KPerBlock / BlockSize;
static_assert(total_pixels % M1 == 0);
constexpr index_t K3 = total_pixels / M1;
constexpr index_t KPack = GetSmemPackA<Problem>();
static_assert(KPack % K3 == 0);
constexpr index_t K2 = KPack / K3;
if constexpr(get_warp_size() % (K2 * M0))
{
constexpr index_t K1 = get_warp_size() / (K2 * M0);
constexpr index_t K0 = BlockSize / get_warp_size();
static_assert(KPerBlock == K0 * K1 * K2 * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
else
{
constexpr index_t K1 = (K2 * M0) / get_warp_size();
constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = BlockSize / get_warp_size() / K1;
static_assert(KPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
}
else
{
constexpr index_t K1 = 16 / sizeof(ADataType);
constexpr index_t K0 = KPerBlock / K1;
constexpr index_t M2 = get_warp_size() / K0;
// coalesce reading for each blocks
if constexpr(get_warp_size() % (M2 * K0) == 0)
{
constexpr index_t M1 = BlockSize / get_warp_size();
static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
constexpr index_t M0 = MPerBlock / (M2 * M1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
else
{
constexpr index_t M0 = BlockSize / get_warp_size();
constexpr index_t M1 = MPerBlock / (M2 * M0);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<0>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
}
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
{
using BDataType = remove_cvref_t<typename Problem::BDataType>;
using BLayout = remove_cvref_t<typename Problem::BLayout>;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
constexpr index_t N1 = Problem::VectorLoadSize / sizeof(BDataType);
constexpr index_t N0 = NPerBlock / N1;
constexpr index_t total_pixels = NPerBlock * KPerBlock / BlockSize;
static_assert(total_pixels % N1 == 0);
constexpr index_t K3 = total_pixels / N1;
constexpr index_t KPack = GetSmemPackB<Problem>();
static_assert(KPack % K3 == 0);
constexpr index_t K2 = KPack / K3;
if constexpr(get_warp_size() % (K2 * N0) == 0)
{
constexpr index_t K1 = get_warp_size() / (K2 * N0);
constexpr index_t K0 = BlockSize / get_warp_size();
static_assert(KPerBlock == K0 * K1 * K2 * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
else
{
constexpr index_t K1 = (K2 * N0) / get_warp_size();
constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = BlockSize / get_warp_size() / K1;
static_assert(KPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
}
else
{
constexpr index_t K1 = Problem::VectorLoadSize / sizeof(BDataType);
constexpr index_t K0 = KPerBlock / K1;
constexpr index_t N2 = get_warp_size() / K0;
// coalesce reading for each blocks
if constexpr(get_warp_size() % (N2 * K0) == 0)
{
constexpr index_t N1 = BlockSize / get_warp_size();
static_assert(N2 != 0, "N2 is zero, which will lead to a division by zero error.");
static_assert(N1 != 0, "N1 is zero, which will lead to a division by zero error.");
constexpr index_t N0 = NPerBlock / (N2 * N1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
// coalesce reading for each warps
else
{
constexpr index_t N0 = BlockSize / get_warp_size();
constexpr index_t N1 = NPerBlock / (N2 * N0);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<0>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
}
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledBRegBlockDescriptor()
{
using BLayout = remove_cvref_t<typename Problem::BLayout>;
using BDataType = remove_cvref_t<typename Problem::BDataType>;
static_assert(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>);
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t K1 = 16 / sizeof(BDataType);
constexpr index_t K0 = kKPerBlock / K1;
constexpr index_t N2 = get_warp_size() / K0;
#if 1 // coalesce reading for each blocks
constexpr index_t N1 = kBlockSize / get_warp_size();
static_assert(N2 != 0, "M2 is zero, which will lead to a division by zero error.");
static_assert(N1 != 0, "M1 is zero, which will lead to a division by zero error.");
constexpr index_t N0 = kNPerBlock / (N2 * N1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
#else // coalesce reading for each warps
constexpr index_t N0 = kBlockSize / get_warp_size();
constexpr index_t N1 = kNPerBlock / (N2 * N0);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<0>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
#endif
constexpr index_t N1 = Problem::VectorLoadSize / sizeof(BDataType);
constexpr index_t N0 = kNPerBlock / N1;
constexpr index_t total_pixels = kNPerBlock * kKPerBlock / kBlockSize;
static_assert(total_pixels % N1 == 0);
constexpr index_t K3 = total_pixels / N1;
constexpr index_t kKPack = GetSmemPackB<Problem>();
static_assert(kKPack % K3 == 0);
constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
constexpr index_t warp_size = get_warp_size();
if constexpr(warp_size % (K2 * N0) == 0)
{
constexpr index_t K1 = warp_size / (K2 * N0);
constexpr index_t K0 = kBlockSize / warp_size;
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
else
{
constexpr index_t K1 = (K2 * N0) / get_warp_size();
constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = kBlockSize / get_warp_size() / K1;
static_assert(kKPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledARegBlockDescriptor()
{
using ALayout = remove_cvref_t<typename Problem::ALayout>;
using ADataType = remove_cvref_t<typename Problem::ADataType>;
static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::RowMajor>);
constexpr index_t kBlockSize = Problem::kBlockSize;
constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t M1 = Problem::VectorLoadSize / sizeof(ADataType);
constexpr index_t M0 = kMPerBlock / M1;
constexpr index_t total_pixels = kMPerBlock * kKPerBlock / kBlockSize;
static_assert(total_pixels % M1 == 0);
constexpr index_t K3 = total_pixels / M1;
constexpr index_t kKPack = GetSmemPackA<Problem>();
static_assert(kKPack % K3 == 0);
constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
constexpr index_t warp_size = get_warp_size();
if constexpr(warp_size % (K2 * M0) == 0)
{
constexpr index_t K1 = warp_size / (K2 * M0);
constexpr index_t K0 = kBlockSize / warp_size;
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
else
{
constexpr index_t K1 = (K2 * M0) / get_warp_size();
constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = kBlockSize / get_warp_size() / K1;
static_assert(kKPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
}
template <typename Problem>
......
include/ck_tile/ops/gemm/pipeline/gemm_pipeline_problem.hpp
View file @ f3bbfe3e
......@@ -3,40 +3,133 @@
#pragma once
#include "ck_tile/core.hpp"
#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
namespace ck_tile {
static constexpr int _VectorSize = 16;
template <typename ADataType_,
typename BDataType_,
typename CDataType_,
typename BlockGemmShape_,
typename TileGemmTraits_>
struct GemmPipelineProblem
struct GemmPipelineProblemBase
{
using ADataType = remove_cvref_t<ADataType_>;
using BDataType = remove_cvref_t<BDataType_>;
using CDataType = remove_cvref_t<CDataType_>;
using GemmTraits = remove_cvref_t<TileGemmTraits_>;
using ADataType = remove_cvref_t<ADataType_>;
using BDataType = remove_cvref_t<BDataType_>;
using CDataType = remove_cvref_t<CDataType_>;
using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
using GemmTraits = remove_cvref_t<TileGemmTraits_>;
using ALayout = remove_cvref_t<typename GemmTraits::ALayout>;
using BLayout = remove_cvref_t<typename GemmTraits::BLayout>;
using CLayout = remove_cvref_t<typename GemmTraits::CLayout>;
static constexpr index_t kBlockSize = BlockGemmShape::NumWarps * get_warp_size();
static constexpr bool kPadA = GemmTraits::kPadA;
static constexpr bool kPadB = GemmTraits::kPadB;
static constexpr bool kPadC = GemmTraits::kPadC;
static constexpr index_t VectorLoadSize = GemmTraits::_VectorSize;
static constexpr index_t kBlockSize = BlockGemmShape::NumWarps * get_warp_size();
static constexpr bool kPadM = GemmTraits::kPadM;
static constexpr bool kPadN = GemmTraits::kPadN;
static constexpr bool kPadK = GemmTraits::kPadK;
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentA()
{
if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
constexpr index_t pixels_per_thread =
BlockGemmShape::kM * BlockGemmShape::kK / kBlockSize;
return pixels_per_thread < VectorLoadSize / sizeof(ADataType)
? pixels_per_thread
: VectorLoadSize / sizeof(ADataType);
}
else
{
return VectorLoadSize / sizeof(ADataType);
}
}
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentB()
{
if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
constexpr index_t pixels_per_thread =
BlockGemmShape::kN * BlockGemmShape::kK / kBlockSize;
return pixels_per_thread < VectorLoadSize / sizeof(BDataType)
? pixels_per_thread
: VectorLoadSize / sizeof(BDataType);
}
else
{
return VectorLoadSize / sizeof(BDataType);
}
}
CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentC()
{
if constexpr(std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
constexpr index_t N1 = kBlockSize / get_warp_size();
constexpr index_t N2 = std::min(BlockGemmShape::kN / N1, get_warp_size());
constexpr index_t M0 = get_warp_size() / N2;
constexpr index_t M1 = BlockGemmShape::kM / M0;
static constexpr index_t VectorSizeA = kPadA ? 1 : _VectorSize / sizeof(ADataType);
static constexpr index_t VectorSizeB = kPadB ? 1 : _VectorSize / sizeof(BDataType);
static constexpr index_t VectorSizeC = kPadC ? 1 : _VectorSize / sizeof(CDataType);
return std::min(M1, static_cast<index_t>(VectorLoadSize / sizeof(CDataType)));
}
else
{
constexpr index_t M1 = kBlockSize / get_warp_size();
constexpr index_t M2 = std::min(BlockGemmShape::kM / M1, get_warp_size());
constexpr index_t N0 = get_warp_size() / M2;
constexpr index_t N1 = BlockGemmShape::kN / N0;
return std::min(N1, static_cast<index_t>(VectorLoadSize / sizeof(CDataType)));
}
}
static constexpr index_t VectorSizeA = []() {
if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
{
return kPadK ? 1 : GetAlignmentA();
}
else
{
return kPadM ? 1 : GetAlignmentA();
}
}();
static constexpr index_t VectorSizeB = []() {
if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
{
return kPadN ? 1 : GetAlignmentB();
}
else
{
return kPadK ? 1 : GetAlignmentB();
}
}();
static constexpr index_t VectorSizeC = []() {
if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
{
return kPadN ? 1 : GetAlignmentC();
}
else
{
return kPadM ? 1 : GetAlignmentC();
}
}();
};
// Alias for GemmPipelineProblem
template <typename ADataType_,
typename BDataType_,
typename CDataType_,
typename BlockGemmShape_,
typename TileGemmTraits_>
using GemmPipelineProblem =
GemmPipelineProblemBase<ADataType_, BDataType_, CDataType_, BlockGemmShape_, TileGemmTraits_>;
template <typename ADataType_,
typename BDataType_,
typename CDataType_,
......@@ -45,30 +138,15 @@ template <typename ADataType_,
GemmPipelineScheduler Scheduler_ = GemmPipelineScheduler::Intrawave,
bool HasHotLoop_ = true,
TailNumber TailNum_ = TailNumber::Full>
struct UniversalGemmPipelineProblem
struct UniversalGemmPipelineProblem : public GemmPipelineProblemBase<ADataType_,
BDataType_,
CDataType_,
BlockGemmShape_,
TileGemmTraits_>
{
using ADataType = remove_cvref_t<ADataType_>;
using BDataType = remove_cvref_t<BDataType_>;
using CDataType = remove_cvref_t<CDataType_>;
using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
using GemmTraits = remove_cvref_t<TileGemmTraits_>;
using ALayout = remove_cvref_t<typename GemmTraits::ALayout>;
using BLayout = remove_cvref_t<typename GemmTraits::BLayout>;
using CLayout = remove_cvref_t<typename GemmTraits::CLayout>;
static constexpr auto Scheduler = Scheduler_;
static constexpr auto HasHotLoop = HasHotLoop_;
static constexpr auto TailNum = TailNum_;
static constexpr index_t kBlockSize = BlockGemmShape::NumWarps * get_warp_size();
static constexpr bool kPadA = GemmTraits::kPadA;
static constexpr bool kPadB = GemmTraits::kPadB;
static constexpr bool kPadC = GemmTraits::kPadC;
static constexpr index_t VectorSizeA = kPadA ? _VectorSize / sizeof(ADataType) : 1;
static constexpr index_t VectorSizeB = kPadB ? _VectorSize / sizeof(BDataType) : 1;
static constexpr index_t VectorSizeC = kPadC ? _VectorSize / sizeof(CDataType) : 1;
static constexpr auto Scheduler = Scheduler_;
static constexpr auto HasHotLoop = HasHotLoop_;
static constexpr auto TailNum = TailNum_;
};
} // namespace ck_tile
include/ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp
View file @ f3bbfe3e
......@@ -9,12 +9,8 @@
namespace ck_tile {
// UniversalGemm Policy
template <typename LayoutA_, typename LayoutB_, typename LayoutC_>
struct UniversalGemmPipelineAgBgCrPolicy
{
using LayoutA = remove_cvref_t<LayoutA_>;
using LayoutB = remove_cvref_t<LayoutB_>;
using LayoutC = remove_cvref_t<LayoutC_>;
static constexpr auto I0 = number<0>{};
static constexpr auto I1 = number<1>{};
......@@ -22,286 +18,136 @@ struct UniversalGemmPipelineAgBgCrPolicy
static constexpr bool TransposeC = true;
template <typename Problem, typename DataType, index_t MNPerBlock>
CK_TILE_HOST_DEVICE static constexpr auto GetVectorLoadSize()
{
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t elements_per_thread = MNPerBlock * KPerBlock / BlockSize;
if constexpr(elements_per_thread % (16 / sizeof(DataType)) == 0)
{
return (16 / sizeof(DataType));
}
else if constexpr(elements_per_thread % (8 / sizeof(DataType)) == 0)
{
return (8 / sizeof(DataType));
}
else if constexpr(elements_per_thread % (4 / sizeof(DataType)) == 0 &&
sizeof(DataType) >= 4)
{
return (4 / sizeof(DataType));
}
else if constexpr(elements_per_thread % (2 / sizeof(DataType)) == 0 &&
sizeof(DataType) >= 2)
{
return (2 / sizeof(DataType));
}
else
{
return 1;
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
{
using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
typename Problem::BDataType,
typename Problem::CDataType,
Problem::BlockGemmShape::WarpTile::at(I0),
Problem::BlockGemmShape::WarpTile::at(I1),
Problem::BlockGemmShape::WarpTile::at(I2),
TransposeC>;
using ADataType = remove_cvref_t<typename Problem::ADataType>;
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t K1 = WarpGemm::kK;
constexpr index_t K0 = KPerBlock / K1;
if constexpr(std::is_same<tensor_layout::gemm::RowMajor, LayoutA>::value)
{
constexpr auto MLdsLayer = 32 * 4 / KPerBlock / sizeof(ADataType) < 1
? 1
: 32 * 4 / KPerBlock / sizeof(ADataType);
constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
make_tuple(K0 * number<MLdsLayer>{}, number<MPerBlock / MLdsLayer>{}, K1),
make_tuple(K1, number<KPerBlock * MLdsLayer>{}, I1));
constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
a_lds_block_desc,
make_tuple(make_xor_transform(make_tuple(number<MPerBlock / MLdsLayer>{},
number<K0 * MLdsLayer>{})),
make_pass_through_transform(K1)),
make_tuple(sequence<1, 0>{}, sequence<2>{}),
make_tuple(sequence<1, 0>{}, sequence<2>{}));
constexpr auto a_lds_block_desc_ak0_kMLdsLayer_m_ak1 = transform_tensor_descriptor(
a_lds_block_desc_permuted,
make_tuple(make_unmerge_transform(make_tuple(K0, number<MLdsLayer>{})),
make_pass_through_transform(number<MPerBlock / MLdsLayer>{}),
make_pass_through_transform(K1)),
make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
constexpr auto a_lds_block_desc_m_k = transform_tensor_descriptor(
a_lds_block_desc_ak0_kMLdsLayer_m_ak1,
make_tuple(make_merge_transform_v3_division_mod(make_tuple(K0, K1)),
make_merge_transform_v3_division_mod(
make_tuple(number<MPerBlock / MLdsLayer>{}, number<MLdsLayer>{}))),
make_tuple(sequence<0, 3>{}, sequence<1, 2>{}),
make_tuple(sequence<1>{}, sequence<0>{}));
return a_lds_block_desc_m_k;
}
else // ColumnMajor A
{
// kfold and mpair dimension is not always required.
// more dimension in merge_transform increase the difficulty of generating immarg offset
// for compiler.
constexpr auto M0 = get_warp_size() * Problem::BlockGemmShape::BlockWarps::at(I0);
constexpr auto M1 = MPerBlock / M0;
constexpr auto KThreadWrite = Problem::kBlockSize / M0;
constexpr auto K0PerThreadWrite = K0 / KThreadWrite;
constexpr auto KThreadRead = 64 / WarpGemm::kM;
constexpr auto K0PerThreadRead = K0 / KThreadRead;
constexpr auto kfold =
(K1 * M0 * sizeof(ADataType) > 128) ? 1 : 128 / (K1 * M0 * sizeof(ADataType));
constexpr auto KThreadReadPerm =
(kfold * K0PerThreadWrite / K0PerThreadRead) > 1
? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
: KThreadRead;
// 1<=mpair<=kN0
constexpr auto mpair = (K1 * WarpGemm::kM * sizeof(ADataType) > 128)
? 1
: ((128 / (K1 * WarpGemm::kM * sizeof(ADataType))) > M0
? M0
: 128 / (K1 * WarpGemm::kM * sizeof(ADataType)));
constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed(
make_tuple(number<KThreadWrite / kfold / KThreadReadPerm>{},
number<K0PerThreadWrite>{},
number<KThreadReadPerm * M1>{},
number<kfold * M0 / mpair>{},
number<mpair>{},
K1));
constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
a_lds_block_desc,
make_tuple(
make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(number<K0PerThreadWrite>{}),
make_xor_transform(
make_tuple(number<KThreadReadPerm * M1>{}, number<kfold * M0 / mpair>{})),
make_pass_through_transform(number<mpair>{}),
make_pass_through_transform(K1)),
make_tuple(
sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}),
make_tuple(
sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}));
constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor(
a_lds_block_desc_permuted,
make_tuple(
make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(number<K0PerThreadWrite>{}),
make_unmerge_transform(make_tuple(number<KThreadReadPerm>{}, number<M1>{})),
make_unmerge_transform(make_tuple(number<kfold>{}, number<M0 / mpair>{})),
make_pass_through_transform(number<mpair>{}),
make_pass_through_transform(K1)),
make_tuple(sequence<0>{},
sequence<1>{},
sequence<2>{},
sequence<3>{},
sequence<4>{},
sequence<5>{}),
make_tuple(sequence<1>{},
sequence<2>{},
sequence<0, 3>{},
sequence<4, 5>{},
sequence<6>{},
sequence<7>{}));
constexpr auto a_lds_block_desc_m_k = transform_tensor_descriptor(
a_lds_block_desc_unmerged,
make_tuple(make_merge_transform_v3_division_mod(
make_tuple(number<KThreadReadPerm>{},
number<KThreadWrite / kfold / KThreadReadPerm>{},
number<kfold>{},
number<K0PerThreadWrite>{},
K1)),
make_merge_transform_v3_division_mod(
make_tuple(number<M0 / mpair>{}, number<mpair>{}, number<M1>{}))),
make_tuple(sequence<0, 1, 4, 2, 7>{}, sequence<5, 6, 3>{}),
make_tuple(sequence<1>{}, sequence<0>{}));
return a_lds_block_desc_m_k;
}
constexpr index_t KPack = GetVectorLoadSize<Problem, ADataType, MPerBlock>();
constexpr auto DataTypeSize = sizeof(ADataType);
constexpr auto MLdsLayer =
(32 * 4 / KPerBlock / DataTypeSize) < 1 ? 1 : (32 * 4 / KPerBlock / DataTypeSize);
constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<KPerBlock / KPack * MLdsLayer>{},
number<MPerBlock / MLdsLayer>{},
number<KPack>{}),
make_tuple(number<KPack>{}, number<KPerBlock * MLdsLayer>{}, number<1>{}),
number<KPack>{},
number<1>{});
constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
a_lds_block_desc_0,
make_tuple(make_xor_transform(make_tuple(number<MPerBlock / MLdsLayer>{},
number<KPerBlock / KPack * MLdsLayer>{})),
make_pass_through_transform(number<KPack>{})),
make_tuple(sequence<1, 0>{}, sequence<2>{}),
make_tuple(sequence<1, 0>{}, sequence<2>{}));
constexpr auto a_lds_block_desc_xk0_mnldslayer_mn_xk1 = transform_tensor_descriptor(
a_lds_block_desc_permuted,
make_tuple(make_unmerge_transform(
make_tuple(number<KPerBlock / KPack>{}, number<MLdsLayer>{})),
make_pass_through_transform(number<MPerBlock / MLdsLayer>{}),
make_pass_through_transform(number<KPack>{})),
make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
constexpr auto a_lds_block_desc = transform_tensor_descriptor(
a_lds_block_desc_xk0_mnldslayer_mn_xk1,
make_tuple(make_merge_transform_v3_division_mod(
make_tuple(number<MPerBlock / MLdsLayer>{}, number<MLdsLayer>{})),
make_merge_transform_v3_division_mod(
make_tuple(number<KPerBlock / KPack>{}, number<KPack>{}))),
make_tuple(sequence<1, 2>{}, sequence<0, 3>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
return a_lds_block_desc;
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
{
using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
typename Problem::BDataType,
typename Problem::CDataType,
Problem::BlockGemmShape::WarpTile::at(I0),
Problem::BlockGemmShape::WarpTile::at(I1),
Problem::BlockGemmShape::WarpTile::at(I2),
TransposeC>;
using BDataType = remove_cvref_t<typename Problem::BDataType>;
constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t K1 = WarpGemm::kK;
constexpr index_t K0 = KPerBlock / K1;
if constexpr(std::is_same<tensor_layout::gemm::ColumnMajor, LayoutB>::value)
{
// NLdsLayer * K0 as logical Bank
constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(BDataType) < 1
? 1
: 32 * 4 / KPerBlock / sizeof(BDataType);
;
constexpr auto b_lds_block_desc = make_naive_tensor_descriptor(
make_tuple(K0 * number<NLdsLayer>{}, number<NPerBlock / NLdsLayer>{}, K1),
make_tuple(K1, number<KPerBlock * NLdsLayer>{}, I1));
constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
b_lds_block_desc,
make_tuple(make_xor_transform(make_tuple(number<NPerBlock / NLdsLayer>{},
number<K0 * NLdsLayer>{})),
make_pass_through_transform(K1)),
make_tuple(sequence<1, 0>{}, sequence<2>{}),
make_tuple(sequence<1, 0>{}, sequence<2>{}));
constexpr auto b_lds_block_desc_bk0_kNLdsLayer_n_bk1 = transform_tensor_descriptor(
b_lds_block_desc_permuted,
make_tuple(make_unmerge_transform(make_tuple(K0, number<NLdsLayer>{})),
make_pass_through_transform(number<NPerBlock / NLdsLayer>{}),
make_pass_through_transform(K1)),
make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
b_lds_block_desc_bk0_kNLdsLayer_n_bk1,
make_tuple(make_merge_transform_v3_division_mod(make_tuple(K0, K1)),
make_merge_transform_v3_division_mod(
make_tuple(number<NPerBlock / NLdsLayer>{}, number<NLdsLayer>{}))),
make_tuple(sequence<0, 3>{}, sequence<1, 2>{}),
make_tuple(sequence<1>{}, sequence<0>{}));
return b_lds_block_desc_n_k;
}
else // RowMajor B
{
constexpr auto N0 = get_warp_size() * Problem::BlockGemmShape::BlockWarps::at(I1);
constexpr auto N1 = NPerBlock / N0;
constexpr auto KThreadWrite = Problem::kBlockSize / N0;
constexpr auto K0PerThreadWrite = K0 / KThreadWrite;
constexpr auto KThreadRead = 64 / WarpGemm::kN;
constexpr auto K0PerThreadRead = K0 / KThreadRead;
constexpr auto kfold =
(K1 * N0 * sizeof(BDataType) > 128) ? 1 : 128 / (K1 * N0 * sizeof(BDataType));
constexpr auto KThreadReadPerm =
(kfold * K0PerThreadWrite / K0PerThreadRead) > 1
? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
: KThreadRead;
// 1<=npair<=kN0
constexpr auto npair = (K1 * WarpGemm::kN * sizeof(BDataType) > 128)
? 1
: ((128 / (K1 * WarpGemm::kN * sizeof(BDataType))) > N0
? N0
: 128 / (K1 * WarpGemm::kN * sizeof(BDataType)));
constexpr auto b_lds_block_desc = make_naive_tensor_descriptor_packed(
make_tuple(number<KThreadWrite / kfold / KThreadReadPerm>{},
number<K0PerThreadWrite>{},
number<KThreadReadPerm * N1>{},
number<kfold * N0 / npair>{},
number<npair>{},
K1));
constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
b_lds_block_desc,
make_tuple(
make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(number<K0PerThreadWrite>{}),
make_xor_transform(
make_tuple(number<KThreadReadPerm * N1>{}, number<kfold * N0 / npair>{})),
make_pass_through_transform(number<npair>{}),
make_pass_through_transform(K1)),
make_tuple(
sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}),
make_tuple(
sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}));
constexpr auto b_lds_block_desc_unmerged = transform_tensor_descriptor(
b_lds_block_desc_permuted,
make_tuple(
make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
make_pass_through_transform(number<K0PerThreadWrite>{}),
make_unmerge_transform(make_tuple(number<KThreadReadPerm>{}, number<N1>{})),
make_unmerge_transform(make_tuple(number<kfold>{}, number<N0 / npair>{})),
make_pass_through_transform(number<npair>{}),
make_pass_through_transform(K1)),
make_tuple(sequence<0>{},
sequence<1>{},
sequence<2>{},
sequence<3>{},
sequence<4>{},
sequence<5>{}),
make_tuple(sequence<1>{},
sequence<2>{},
sequence<0, 3>{},
sequence<4, 5>{},
sequence<6>{},
sequence<7>{}));
constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
b_lds_block_desc_unmerged,
make_tuple(make_merge_transform_v3_division_mod(
make_tuple(number<KThreadReadPerm>{},
number<KThreadWrite / kfold / KThreadReadPerm>{},
number<kfold>{},
number<K0PerThreadWrite>{},
K1)),
make_merge_transform_v3_division_mod(
make_tuple(number<N0 / npair>{}, number<npair>{}, number<N1>{}))),
make_tuple(sequence<0, 1, 4, 2, 7>{}, sequence<5, 6, 3>{}),
make_tuple(sequence<1>{}, sequence<0>{}));
return b_lds_block_desc_n_k;
}
constexpr index_t KPack = GetVectorLoadSize<Problem, BDataType, NPerBlock>();
constexpr auto DataTypeSize = sizeof(BDataType);
constexpr auto NLdsLayer =
(32 * 4 / KPerBlock / DataTypeSize) < 1 ? 1 : (32 * 4 / KPerBlock / DataTypeSize);
constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
make_tuple(number<KPerBlock / KPack * NLdsLayer>{},
number<NPerBlock / NLdsLayer>{},
number<KPack>{}),
make_tuple(number<KPack>{}, number<KPerBlock * NLdsLayer>{}, number<1>{}),
number<KPack>{},
number<1>{});
constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
b_lds_block_desc_0,
make_tuple(make_xor_transform(make_tuple(number<NPerBlock / NLdsLayer>{},
number<KPerBlock / KPack * NLdsLayer>{})),
make_pass_through_transform(number<KPack>{})),
make_tuple(sequence<1, 0>{}, sequence<2>{}),
make_tuple(sequence<1, 0>{}, sequence<2>{}));
constexpr auto b_lds_block_desc_xk0_mnldslayer_mn_xk1 = transform_tensor_descriptor(
b_lds_block_desc_permuted,
make_tuple(make_unmerge_transform(
make_tuple(number<KPerBlock / KPack>{}, number<NLdsLayer>{})),
make_pass_through_transform(number<NPerBlock / NLdsLayer>{}),
make_pass_through_transform(number<KPack>{})),
make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
constexpr auto b_lds_block_desc = transform_tensor_descriptor(
b_lds_block_desc_xk0_mnldslayer_mn_xk1,
make_tuple(make_merge_transform_v3_division_mod(
make_tuple(number<NPerBlock / NLdsLayer>{}, number<NLdsLayer>{})),
make_merge_transform_v3_division_mod(
make_tuple(number<KPerBlock / KPack>{}, number<KPack>{}))),
make_tuple(sequence<1, 2>{}, sequence<0, 3>{}),
make_tuple(sequence<0>{}, sequence<1>{}));
return b_lds_block_desc;
}
template <typename Problem>
......@@ -334,69 +180,268 @@ struct UniversalGemmPipelineAgBgCrPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
{
using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
typename Problem::BDataType,
typename Problem::CDataType,
Problem::BlockGemmShape::WarpTile::at(I0),
Problem::BlockGemmShape::WarpTile::at(I1),
Problem::BlockGemmShape::WarpTile::at(I2),
TransposeC>;
using ADataType = remove_cvref_t<typename Problem::ADataType>;
using ALayout = remove_cvref_t<typename Problem::ALayout>;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t K1 = WarpGemm::kK;
constexpr index_t K0 = KPerBlock / K1;
constexpr index_t M2 = get_warp_size() / K0;
constexpr index_t M1 = BlockSize / get_warp_size();
static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
constexpr index_t M0 = MPerBlock / (M2 * M1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
{
constexpr index_t M1 = Problem::VectorLoadSize / sizeof(ADataType);
constexpr index_t M0 = MPerBlock / M1;
constexpr index_t total_pixels = MPerBlock * KPerBlock / BlockSize;
static_assert(total_pixels % M1 == 0);
constexpr index_t K3 = total_pixels / M1;
constexpr index_t KPack = GetVectorLoadSize<Problem, ADataType, MPerBlock>();
static_assert(KPack % K3 == 0);
constexpr index_t K2 = KPack / K3;
if constexpr(get_warp_size() % (K2 * M0) == 0)
{
constexpr index_t K1 = get_warp_size() / (K2 * M0);
constexpr index_t K0 = BlockSize / get_warp_size();
static_assert(KPerBlock == K0 * K1 * K2 * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
else
{
constexpr index_t K1 = (K2 * M0) / get_warp_size();
constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = BlockSize / get_warp_size() / K1;
static_assert(KPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
}
else
{
constexpr index_t K1 = Problem::VectorLoadSize / sizeof(ADataType);
constexpr index_t K0 = KPerBlock / K1;
constexpr index_t M2 = get_warp_size() / K0;
if constexpr(get_warp_size() % (M2 * K0) == 0)
{
constexpr index_t M1 = BlockSize / get_warp_size();
static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
constexpr index_t M0 = MPerBlock / (M2 * M1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
else
{
constexpr index_t M0 = BlockSize / get_warp_size();
constexpr index_t M1 = MPerBlock / (M2 * M0);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<0>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
}
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
{
using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
typename Problem::BDataType,
typename Problem::CDataType,
Problem::BlockGemmShape::WarpTile::at(I0),
Problem::BlockGemmShape::WarpTile::at(I1),
Problem::BlockGemmShape::WarpTile::at(I2),
TransposeC>;
using BDataType = remove_cvref_t<typename Problem::BDataType>;
using BLayout = remove_cvref_t<typename Problem::BLayout>;
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
{
constexpr index_t N1 = Problem::VectorLoadSize / sizeof(BDataType);
constexpr index_t N0 = NPerBlock / N1;
constexpr index_t total_pixels = NPerBlock * KPerBlock / BlockSize;
static_assert(total_pixels % N1 == 0);
constexpr index_t K3 = total_pixels / N1;
constexpr index_t KPack = GetVectorLoadSize<Problem, BDataType, NPerBlock>();
static_assert(KPack % K3 == 0);
constexpr index_t K2 = KPack / K3;
if constexpr(get_warp_size() % (K2 * N0) == 0)
{
constexpr index_t K1 = get_warp_size() / (K2 * N0);
constexpr index_t K0 = BlockSize / get_warp_size();
static_assert(KPerBlock == K0 * K1 * K2 * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
else
{
constexpr index_t K1 = (K2 * N0) / get_warp_size();
constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = BlockSize / get_warp_size() / K1;
static_assert(KPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<2, 1>,
sequence<3, 1>>{});
}
}
else
{
constexpr index_t K1 = Problem::VectorLoadSize / sizeof(BDataType);
constexpr index_t K0 = KPerBlock / K1;
constexpr index_t N2 = get_warp_size() / K0;
// coalesce reading for each blocks
if constexpr(get_warp_size() % (N2 * K0) == 0)
{
constexpr index_t N1 = BlockSize / get_warp_size();
static_assert(N2 != 0, "N2 is zero, which will lead to a division by zero error.");
static_assert(N1 != 0, "N1 is zero, which will lead to a division by zero error.");
constexpr index_t N0 = NPerBlock / (N2 * N1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
}
// coalesce reading for each warps
else
{
constexpr index_t N0 = BlockSize / get_warp_size();
constexpr index_t N1 = NPerBlock / (N2 * N0);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<0>, sequence<2, 0>>,
sequence<1, 2>,
sequence<1, 1>>{});
}
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledARegBlockDescriptor()
{
using ALayout = remove_cvref_t<typename Problem::ALayout>;
using ADataType = remove_cvref_t<typename Problem::ADataType>;
static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t MPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t M1 = Problem::VectorLoadSize / sizeof(ADataType);
constexpr index_t M0 = MPerBlock / M1;
constexpr index_t total_pixels = MPerBlock * KPerBlock / BlockSize;
static_assert(total_pixels % M1 == 0);
constexpr index_t K3 = total_pixels / M1;
constexpr index_t kKPack = GetVectorLoadSize<Problem, ADataType, MPerBlock>();
static_assert(kKPack % K3 == 0);
constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
constexpr index_t warp_size = get_warp_size();
if constexpr(warp_size % (K2 * M0) == 0)
{
constexpr index_t K1 = warp_size / (K2 * M0);
constexpr index_t K0 = BlockSize / warp_size;
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
else
{
constexpr index_t K1 = (K2 * M0) / get_warp_size();
constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = BlockSize / get_warp_size() / K1;
static_assert(KPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledBRegBlockDescriptor()
{
using BLayout = remove_cvref_t<typename Problem::BLayout>;
using BDataType = remove_cvref_t<typename Problem::BDataType>;
static_assert(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>);
constexpr index_t BlockSize = Problem::kBlockSize;
constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
constexpr index_t K1 = WarpGemm::kK;
constexpr index_t K0 = KPerBlock / K1;
constexpr index_t N2 = get_warp_size() / K0;
constexpr index_t N1 = BlockSize / get_warp_size();
static_assert(N2 != 0, "M2 is zero, which will lead to a division by zero error.");
static_assert(N1 != 0, "M1 is zero, which will lead to a division by zero error.");
constexpr index_t N0 = NPerBlock / (N2 * N1);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
tuple<sequence<1>, sequence<1, 2>>,
tuple<sequence<1>, sequence<2, 0>>,
sequence<1, 2>,
sequence<0, 1>>{});
constexpr index_t N1 = Problem::VectorLoadSize / sizeof(BDataType);
constexpr index_t N0 = NPerBlock / N1;
constexpr index_t total_pixels = NPerBlock * KPerBlock / BlockSize;
static_assert(total_pixels % N1 == 0);
constexpr index_t K3 = total_pixels / N1;
constexpr index_t kKPack = GetVectorLoadSize<Problem, BDataType, NPerBlock>();
static_assert(kKPack % K3 == 0);
constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
constexpr index_t warp_size = get_warp_size();
if constexpr(warp_size % (K2 * N0) == 0)
{
constexpr index_t K1 = warp_size / (K2 * N0);
constexpr index_t K0 = BlockSize / warp_size;
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2, K3>>,
tuple<sequence<2>, sequence<2, 1, 2>>,
tuple<sequence<0>, sequence<1, 0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
else
{
constexpr index_t K1 = (K2 * N0) / get_warp_size();
constexpr index_t K2_m = K2 / K1;
constexpr index_t K0 = BlockSize / get_warp_size() / K1;
static_assert(KPerBlock == K0 * K1 * K2_m * K3);
return make_static_tile_distribution(
tile_distribution_encoding<sequence<1>,
tuple<sequence<N0, N1>, sequence<K0, K1, K2_m, K3>>,
tuple<sequence<2, 2>, sequence<1, 2>>,
tuple<sequence<0, 1>, sequence<0, 2>>,
sequence<1, 2>,
sequence<1, 3>>{});
}
}
template <typename Problem>
......
include/ck_tile/ops/gemm/pipeline/tile_gemm_traits.hpp
View file @ f3bbfe3e
......@@ -3,19 +3,23 @@
#pragma once
#include "ck_tile/core.hpp"
namespace ck_tile {
template <bool kPadA_,
bool kPadB_,
bool kPadC_,
template <bool kPadM_,
bool kPadN_,
bool kPadK_,
typename ALayout_,
typename BLayout_,
typename CLayout_>
struct TileGemmTraits
{
static constexpr bool kPadA = kPadA_;
static constexpr bool kPadB = kPadB_;
static constexpr bool kPadC = kPadC_;
static constexpr bool kPadM = kPadM_;
static constexpr bool kPadN = kPadN_;
static constexpr bool kPadK = kPadK_;
static constexpr int _VectorSize = 16;
using ALayout = ALayout_;
using BLayout = BLayout_;
......
include/ck_tile/ops/layernorm2d/kernel/layernorm2d_fwd_kernel.hpp
View file @ f3bbfe3e
......@@ -28,7 +28,10 @@ struct Layernorm2dFwdHostArgs
index_t m;
index_t n;
index_t stride; // row_stride
index_t x_stride; // x row_stride
index_t xr_stride; // x residule row stride
index_t y_stride; // y row stride
index_t yr_stride; // y residule row stride
};
// TODO: Extract some type to wrapper class
......@@ -93,7 +96,10 @@ struct Layernorm2dFwd
index_t m;
index_t n;
index_t stride; // row_stride
index_t x_stride; // x row_stride
index_t xr_stride; // x residule row stride
index_t y_stride; // y row stride
index_t yr_stride; // y residule row stride
};
using Hargs = Layernorm2dFwdHostArgs;
......@@ -112,7 +118,10 @@ struct Layernorm2dFwd
hargs.epsilon,
hargs.m,
hargs.n,
hargs.stride};
hargs.x_stride,
hargs.xr_stride,
hargs.y_stride,
hargs.yr_stride};
}
CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
......@@ -182,7 +191,7 @@ struct Layernorm2dFwd
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const XDataType*>(kargs.p_x),
make_tuple(kargs.m, kargs.n),
make_tuple(kargs.stride, 1),
make_tuple(kargs.x_stride, 1),
number<Vector_N>{},
number<1>{});
......@@ -201,7 +210,7 @@ struct Layernorm2dFwd
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const XResidualDataType*>(kargs.p_x_residual),
make_tuple(kargs.m, kargs.n),
make_tuple(kargs.stride, 1),
make_tuple(kargs.xr_stride, 1),
number<Vector_N>{},
number<1>{});
......@@ -250,7 +259,7 @@ struct Layernorm2dFwd
auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<YDataType*>(kargs.p_y),
make_tuple(kargs.m, kargs.n),
make_tuple(kargs.stride, 1),
make_tuple(kargs.y_stride, 1),
number<Vector_N>{},
number<1>{});
......@@ -266,7 +275,7 @@ struct Layernorm2dFwd
auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<YResidualDataType*>(kargs.p_y_residual),
make_tuple(kargs.m, kargs.n),
make_tuple(kargs.stride, 1),
make_tuple(kargs.yr_stride, 1),
number<Vector_N>{},
number<1>{});
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_default_policy.hpp
View file @ f3bbfe3e
......@@ -47,7 +47,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
{
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
typename Problem::BlockShape,
Problem::Traits::kFastFDiv>;
return BlockWelford<P_>{};
}
......@@ -57,7 +58,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
{
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
typename Problem::BlockShape,
Problem::Traits::kFastFDiv>;
return BlockWelfordSync<P_>{};
}
......@@ -67,7 +69,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
{
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
typename Problem::BlockShape,
Problem::Traits::kFastFDiv>;
return BlockWelfordCrossWarpSync<P_>{};
}
......@@ -79,7 +82,8 @@ struct Layernorm2dFwdPipelineDefaultPolicy
{
using P_ = BlockWelfordProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
typename Problem::BlockShape,
Problem::Traits::kFastFDiv>;
using block_welford = BlockWelford<P_>;
using x_block_tile =
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_one_pass.hpp
View file @ f3bbfe3e
......@@ -36,6 +36,7 @@ struct Layernorm2dFwdPipelineOnePass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr bool kFastFDiv = Problem::Traits::kFastFDiv;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
......@@ -120,12 +121,20 @@ struct Layernorm2dFwdPipelineOnePass
auto [mean, var] = block_welford(acc, cur_count, max_count);
block_welford_sync(mean, var, cur_count);
block_welford_cross_warp_sync(mean, var, cur_count, smem);
block_tile_welford_post_scale_var(var, cur_count);
block_tile_welford_post_scale_var(var, cur_count, constant<kFastFDiv>{});
// compute inv-std
auto inv_std = tile_elementwise_in(
[&](const auto& v_) {
return type_convert<ComputeDataType>(1.0f) / (sqrt(v_ + epsilon));
if(kFastFDiv && std::is_same_v<ComputeDataType, float>)
{
return type_convert<ComputeDataType>(1.0f) *
__builtin_amdgcn_rcpf(sqrt(v_ + epsilon));
}
else
{
return type_convert<ComputeDataType>(1.0f) / sqrt(v_ + epsilon);
}
},
var);
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_two_pass.hpp
View file @ f3bbfe3e
......@@ -35,6 +35,7 @@ struct Layernorm2dFwdPipelineTwoPass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr bool kFastFDiv = Problem::Traits::kFastFDiv;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
......@@ -137,15 +138,22 @@ struct Layernorm2dFwdPipelineTwoPass
block_welford_sync(mean, var, cur_count);
block_welford_cross_warp_sync(mean, var, cur_count, smem);
block_tile_welford_post_scale_var(var, cur_count);
block_tile_welford_post_scale_var(var, cur_count, constant<kFastFDiv>{});
// compute inv-std
auto inv_std = tile_elementwise_in(
[&](const auto& v_) {
return type_convert<ComputeDataType>(1.0f) / (sqrt(v_ + epsilon));
if(kFastFDiv && std::is_same_v<ComputeDataType, float>)
{
return type_convert<ComputeDataType>(1.0f) *
__builtin_amdgcn_rcpf(sqrt(v_ + epsilon));
}
else
{
return type_convert<ComputeDataType>(1.0f) / sqrt(v_ + epsilon);
}
},
var);
if constexpr(kSaveMean)
store_tile(mean_window, cast_tile<MeanDataType>(mean));
if constexpr(kSaveInvStd)
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_traits.hpp
View file @ f3bbfe3e
......@@ -39,6 +39,7 @@ template<> struct Layernorm2dFusedQuantEnumName<Layernorm2dFusedQuantEnum::SMOOT
template <bool kPadN_,
bool kSaveMeanInvStd_,
bool kFastFDiv_,
bool kTwoPass_,
Layernorm2dFusedAddEnum kFusedAdd_,
Layernorm2dFusedQuantEnum kFusedQuant_>
......@@ -46,6 +47,7 @@ struct Layernorm2dFwdTraits
{
static constexpr bool kPadN = kPadN_;
static constexpr bool kSaveMeanInvStd = kSaveMeanInvStd_;
static constexpr bool kFastFDiv = kFastFDiv_;
static constexpr bool kTwoPass = kTwoPass_;
static constexpr Layernorm2dFusedAddEnum kFusedAdd = kFusedAdd_;
static constexpr Layernorm2dFusedQuantEnum kFusedQuant = kFusedQuant_;
......
include/ck_tile/ops/moe_sorting.hpp 0 → 100644
View file @ f3bbfe3e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp"
#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp"
#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp"
#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_problem.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
include/ck_tile/ops/welford/block/block_welford.hpp
View file @ f3bbfe3e
......@@ -11,9 +11,10 @@ namespace ck_tile {
template <typename Problem_, typename Policy_ = void>
struct BlockWelford
{
using Problem = remove_cvref_t<Problem_>;
using XDataType = typename Problem::XDataType;
using ComputeDataType = typename Problem::ComputeDataType;
using Problem = remove_cvref_t<Problem_>;
using XDataType = typename Problem::XDataType;
using ComputeDataType = typename Problem::ComputeDataType;
static constexpr bool kFastFDiv = Problem::kFastFDiv;
CK_TILE_DEVICE constexpr BlockWelford() {}
......@@ -46,8 +47,11 @@ struct BlockWelford
auto x = ck_tile::type_convert<ComputeDataType>(x_tensor[in_dstr_idx]);
welford_update(
mean_tensor(out_dstr_idx), var_tensor(out_dstr_idx), x, cur_count_);
welford_update(mean_tensor(out_dstr_idx),
var_tensor(out_dstr_idx),
x,
cur_count_,
constant<kFastFDiv>{});
});
}
});
......@@ -89,7 +93,8 @@ struct BlockWelford
template <typename Problem_, typename Policy_ = void>
struct BlockWelfordSync
{
using Problem = remove_cvref_t<Problem_>;
using Problem = remove_cvref_t<Problem_>;
static constexpr bool kFastFDiv = Problem::kFastFDiv;
template <typename MeanDistributedTensor_, typename VarDistributedTensor_>
CK_TILE_DEVICE void
......@@ -157,7 +162,8 @@ struct BlockWelfordSync
v_local_count,
v_remote_mean,
v_remote_var,
v_remote_count);
v_remote_count,
constant<kFastFDiv>{});
});
}
});
......@@ -173,8 +179,9 @@ struct BlockWelfordSync
template <typename Problem_, typename Policy_ = void>
struct BlockWelfordCrossWarpSync
{
using Problem = remove_cvref_t<Problem_>;
using BlockShape = typename Problem::BlockShape;
using Problem = remove_cvref_t<Problem_>;
using BlockShape = typename Problem::BlockShape;
static constexpr bool kFastFDiv = Problem::kFastFDiv;
template <typename MeanDistributedTensor_>
CK_TILE_DEVICE static constexpr index_t GetReduceWarps()
......@@ -304,7 +311,8 @@ struct BlockWelfordCrossWarpSync
v_local_count,
v_remote_mean,
v_remote_var,
v_remote_count);
v_remote_count,
constant<kFastFDiv>{});
});
mean_tensor.get_thread_buffer()(i_0) = v_local_mean;
......@@ -351,12 +359,23 @@ CK_TILE_DEVICE constexpr index_t block_tile_welford_calculate_max_count(int row_
}
// Note: this function must be called after all the computation
template <typename VarDistributedTensor_>
template <typename VarDistributedTensor_, bool FastFdiv_ = false>
CK_TILE_DEVICE constexpr void block_tile_welford_post_scale_var(VarDistributedTensor_& var_tensor,
int count)
int count,
bool_constant<FastFdiv_> = {})
{
using DataType = typename VarDistributedTensor_::DataType;
tile_elementwise_inout([&count](auto& x) { x = x / type_convert<DataType>(count); },
var_tensor);
tile_elementwise_inout(
[&count](auto& x) {
if(FastFdiv_ && std::is_same_v<DataType, float>)
{
x = x * __builtin_amdgcn_rcpf(type_convert<DataType>(count));
}
else
{
x = x / type_convert<DataType>(count);
}
},
var_tensor);
}
} // namespace ck_tile
include/ck_tile/ops/welford/block/block_welford_problem.hpp
View file @ f3bbfe3e
......@@ -7,12 +7,13 @@
namespace ck_tile {
template <typename XDataType_, typename ComputeDataType_, typename BlockShape_>
template <typename XDataType_, typename ComputeDataType_, typename BlockShape_, bool kFastFDiv_>
struct BlockWelfordProblem
{
using XDataType = remove_cvref_t<XDataType_>;
using ComputeDataType = remove_cvref_t<ComputeDataType_>;
using BlockShape = remove_cvref_t<BlockShape_>;
using XDataType = remove_cvref_t<XDataType_>;
using ComputeDataType = remove_cvref_t<ComputeDataType_>;
using BlockShape = remove_cvref_t<BlockShape_>;
static constexpr bool kFastFDiv = kFastFDiv_;
};
} // namespace ck_tile
include/ck_tile/ops/welford/thread/thread_welford.hpp
View file @ f3bbfe3e
......@@ -7,25 +7,46 @@
namespace ck_tile {
template <typename T>
CK_TILE_DEVICE void welford_update(T& mean, T& var, T x, int count)
template <typename T, bool kFastFDiv = false>
CK_TILE_DEVICE void welford_update(T& mean, T& var, T x, int count, bool_constant<kFastFDiv> = {})
{
// TODO: check nan? maybe no
T delta = x - mean;
mean += delta / count;
if(kFastFDiv && std::is_same_v<T, float>)
{
mean += delta * __builtin_amdgcn_rcpf(count);
}
else
{
mean += delta / count;
}
T delta2 = x - mean;
var += delta * delta2;
}
template <typename T>
CK_TILE_DEVICE static void
welford_merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b)
template <typename T, bool kFastFDiv = false>
CK_TILE_DEVICE static void welford_merge(T& mean_a,
T& var_a,
int& count_a,
T mean_b,
T var_b,
int count_b,
bool_constant<kFastFDiv> = {})
{
int count = count_a + count_b;
T count_ = type_convert<T>(count);
T count_a_ = type_convert<T>(count_a);
T count_b_ = type_convert<T>(count_b);
T count_b_over_count = count == 0 ? type_convert<T>(0) : count_b_ / count_;
int count = count_a + count_b;
T count_ = type_convert<T>(count);
T count_a_ = type_convert<T>(count_a);
T count_b_ = type_convert<T>(count_b);
T count_b_over_count;
if(kFastFDiv && std::is_same_v<T, float>)
{
count_b_over_count =
count == 0 ? type_convert<T>(0) : count_b_ * __builtin_amdgcn_rcpf(count_);
}
else
{
count_b_over_count = count == 0 ? type_convert<T>(0) : count_b_ / count_;
}
T delta = mean_b - mean_a;
mean_a += delta * count_b_over_count;
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp
View file @ f3bbfe3e
......@@ -39,7 +39,25 @@ template <ck::index_t NDimSpatial,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_f16_instances = std::tuple<
using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_generic_instances =
std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
......@@ -64,7 +82,25 @@ template <ck::index_t NDimSpatial,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances = std::tuple<
using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_generic_instances =
std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
......@@ -82,6 +118,24 @@ using device_grouped_conv_bwd_weight_two_stage_xdl_c_shuffle_bf16_instances = st
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_generic_instances =
std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>
// clang-format on
>;
// NGCHW requires transpose, we use vector loads and stores params for them
template <ck::index_t NDimSpatial,
typename ALayout,
......@@ -122,6 +176,24 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename ELayout,
ConvolutionBackwardWeightSpecialization ConvSpec,
BlockGemmPipelineScheduler Scheduler,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_generic_instances =
std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransfer| CBlockTransfer| BlockGemm| BlockGemm| NumGroups|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| |
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | |
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>
// clang-format on
>;
template <ck::index_t NDimSpatial,
typename ALayout,
typename BLayout,
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
View file @ f3bbfe3e
......@@ -352,6 +352,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev5_instances(
......@@ -375,6 +377,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev5_instances(
......@@ -390,6 +394,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>)
{
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev2_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_f16_pipev5_instances(
......@@ -403,6 +409,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<ComputeTypeA, ck::bhalf_t> &&
is_same_v<ComputeTypeB, ck::bhalf_t>)
{
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev2_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_xdl_ngchw_gkyxc_ngkhw_bf16_pipev5_instances(
......@@ -464,6 +472,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev5_instances(
......@@ -487,6 +497,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev5_instances(
......@@ -511,6 +523,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>)
{
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev2_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_f16_pipev5_instances(
......@@ -524,6 +538,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<ComputeTypeA, ck::bhalf_t> &&
is_same_v<ComputeTypeB, ck::bhalf_t>)
{
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev2_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ngcdhw_gkzyxc_ngkdhw_bf16_pipev5_instances(
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment